aboutsummaryrefslogtreecommitdiffstats
path: root/backfill
diff options
context:
space:
mode:
authorBryan Newbold <bnewbold@archive.org>2018-03-30 18:29:29 -0700
committerBryan Newbold <bnewbold@archive.org>2018-03-30 18:29:29 -0700
commitf6915b4b44e312cee7eda9626d0330268ab786e2 (patch)
treed4c3436d1bf010a269ab612d23e39e8ae2e9a52d /backfill
parent1b6d04fffb6554fd8ef1479b29c7154df07e4d21 (diff)
downloadsandcrawler-f6915b4b44e312cee7eda9626d0330268ab786e2.tar.gz
sandcrawler-f6915b4b44e312cee7eda9626d0330268ab786e2.zip
refactor backfill for mrjob
Diffstat (limited to 'backfill')
-rw-r--r--backfill/README.md8
-rwxr-xr-xbackfill/backfill_hbase_from_cdx.py126
-rw-r--r--backfill/tests/files/example.cdx20
-rw-r--r--backfill/tests/test_backfill_hbase_from_cdx.py55
4 files changed, 145 insertions, 64 deletions
diff --git a/backfill/README.md b/backfill/README.md
new file mode 100644
index 0000000..90b4ba7
--- /dev/null
+++ b/backfill/README.md
@@ -0,0 +1,8 @@
+
+Run tests:
+
+ pipenv run python -m pytest
+
+Run locally on a file:
+
+ ./backfill_hbase_from_cdx.py tests/files/example.cdx
diff --git a/backfill/backfill_hbase_from_cdx.py b/backfill/backfill_hbase_from_cdx.py
index 757794a..92a6d32 100755
--- a/backfill/backfill_hbase_from_cdx.py
+++ b/backfill/backfill_hbase_from_cdx.py
@@ -6,6 +6,7 @@ formats.
Requires:
- happybase
+- mrjob
TODO:
- argparse
@@ -18,6 +19,8 @@ TODO:
import sys
import json
import happybase
+import mrjob
+from mrjob.job import MRJob
NORMAL_MIME = (
'application/pdf',
@@ -90,80 +93,75 @@ def test_transform_line():
assert transform_line(raw + " extra_field") == correct
-def run(in_lines, out_lines, status_lines, table, mime_filter=None):
+class MRCDXBackfillHBase(MRJob):
- if mime_filter is None:
- mime_filter = ['application/pdf']
- count_skip = count_invalid = count_fail = count_success = 0
+ # CDX lines in
+ INPUT_PROTOCOL = mrjob.protocol.RawValueProtocol
+ OUTPUT_PROTOCOL = mrjob.protocol.JSONValueProtocol
+
+ def configure_args(self):
+ super(MRCDXBackfillHBase, self).configure_args()
+
+ self.add_passthru_arg('--hbase-table',
+ type=str,
+ default='wbgrp-journal-extract-0-qa',
+ help='HBase table to backfill into (must exist)')
+ self.add_passthru_arg('--hbase-host',
+ type=str,
+ default='localhost',
+ help='HBase thrift API host to connect to')
+
+ def __init__(self, *args, **kwargs):
+
+ # Allow passthrough for tests
+ if 'hb_table' in kwargs:
+ self.hb_table = kwargs.pop('hb_table')
+ else:
+ self.hb_table = None
+
+ super(MRCDXBackfillHBase, self).__init__(*args, **kwargs)
+
+ def mapper_init(self):
+
+ if self.hb_table is None:
+ try:
+ host = self.options.hbase_host
+ hb_conn = happybase.Connection(host=host)
+ except Exception as err:
+ raise Exception("Couldn't connect to HBase using host: {}".format(host))
+ self.hb_table = hb_conn.table(self.options.hbase_table)
+
+ self.mime_filter = ['application/pdf']
+
+ def mapper(self, _, raw_cdx):
+
+ self.increment_counter('lines', 'total')
- for raw_cdx in in_lines.readlines():
if (raw_cdx.startswith(' ') or raw_cdx.startswith('filedesc') or
raw_cdx.startswith('#')):
+
# Skip line
- count_skip += 1
- continue
+ self.increment_counter('lines', 'invalid')
+ return _, status
info = transform_line(raw_cdx)
if info is None:
- count_invalid += 1
- continue
- if info['file:mime'] not in mime_filter:
- count_skip += 1
- continue
+ self.increment_counter('lines', 'invalid')
+ return
+
+ if info['file:mime'] not in self.mime_filter:
+ self.increment_counter('lines', 'skip')
+ return
key = info.pop('key')
info['file:cdx'] = json.dumps(info['file:cdx'], sort_keys=True,
indent=None)
- try:
- table.put(key, info)
- count_success += 1
- except:
- status_lines.write("ERROR\n") # TODO:
- count_fail += 1
-
- status_lines.write('\n')
- status_lines.write('skip\t{}\n'.format(count_skip))
- status_lines.write('invalid\t{}\n'.format(count_invalid))
- status_lines.write('fail\t{}\n'.format(count_fail))
- status_lines.write('success\t{}\n'.format(count_success))
-
-
-def test_run():
-
- import io
- import happybase_mock
-
- out = io.StringIO()
- status = io.StringIO()
- raw = io.StringIO("""
-com,sagepub,cep)/content/28/9/960.full.pdf 20170705062200 http://cep.sagepub.com/content/28/9/960.full.pdf application/pdf 301 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ - - 401 313356621 CITESEERX-CRAWL-2017-06-20-20170705061647307-00039-00048-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705062052659-00043-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
-eu,eui,cadmus)/bitstream/handle/1814/36635/rscas_2015_03.pdf;jsessionid=761393014319a39f40d32ae3eb3a853f?sequence=1 20170705062202 http://cadmus.eui.eu/bitstream/handle/1814/36635/RSCAS_2015_03.pdf%3Bjsessionid%3D761393014319A39F40D32AE3EB3A853F?sequence%3D1 application/PDF 200 MPCXVWMUTRUGFP36SLPHKDLY6NGU4S3J - - 854156 328850624 CITESEERX-CRAWL-2017-06-20-20170705061647307-00039-00048-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705062052659-00043-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
-com,pbworks,educ333b)/robots.txt 20170705063311 http://educ333b.pbworks.com/robots.txt text/plain 200 6VAUYENMOU2SK2OWNRPDD6WTQTECGZAD - - 638 398190140 CITESEERX-CRAWL-2017-06-20-20170705062707827-00049-00058-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705063158203-00053-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
-""")
-
- conn = happybase_mock.Connection()
- conn.create_table('wbgrp-journal-extract-test', {'file': {}, 'grobid0': {}})
-
- table = conn.table('wbgrp-journal-extract-test')
- run(raw, out, status, table)
-
- print(status.getvalue())
-
- assert table.row(b'1') == {}
- # HTTP 301
- assert table.row(b'3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ') == {}
- # valid
- assert table.row(b'MPCXVWMUTRUGFP36SLPHKDLY6NGU4S3J') != {}
- # text/plain
- assert table.row(b'6VAUYENMOU2SK2OWNRPDD6WTQTECGZAD') == {}
-
- row = table.row(b'MPCXVWMUTRUGFP36SLPHKDLY6NGU4S3J')
- assert row[b'file:mime'] == b"application/pdf"
- json.loads(row[b'file:cdx'].decode('utf-8'))
-
-if __name__=="__main__":
- hb = happybase.Connection(host='')
- with hb.connection() as conn:
- table = conn.table('wbgrp-journal-extract-0-qa')
- run(sys.stdin, sys.stdout, sys.stderr, table)
+
+ self.hb_table.put(key, info)
+ self.increment_counter('lines', 'success')
+
+ yield _, dict(status="success")
+
+if __name__ == '__main__':
+ MRCDXBackfillHBase.run()
diff --git a/backfill/tests/files/example.cdx b/backfill/tests/files/example.cdx
new file mode 100644
index 0000000..84e3271
--- /dev/null
+++ b/backfill/tests/files/example.cdx
@@ -0,0 +1,20 @@
+edu,cmu,cs,adm,reports-archive)/anon/usr0/ftp/usr0/anon/2002/cmu-cs-02-119.pdf 20170706005950 http://reports-archive.adm.cs.cmu.edu/anon/usr0/ftp/usr0/anon/2002/CMU-CS-02-119.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 361006 17120058 CITESEERX-CRAWL-2017-06-20-20170706004100259-00924-00932-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170706005946792-00926-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+fi,tkk,lib)/diss/2001/isbn951225459x/isbn951225459x.pdf 20170705074926 http://lib.tkk.fi/Diss/2001/isbn951225459X/isbn951225459X.pdf application/pdf 200 KJBCOT7LGBNIAVGEGPUELK5OK6RTFORR - - 344175 255650124 CITESEERX-CRAWL-2017-06-20-20170705074433815-00129-00138-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705074843696-00134-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+org,oxfordjournals,nar)/cgi/reprint/gkl1060v1.pdf 20170706035441 http://nar.oxfordjournals.org/cgi/reprint/gkl1060v1.pdf text/html 301 OX6MLVDFURLT2KSYCXUYW2PZNOVFSEVF - - 697 49346051 CITESEERX-CRAWL-2017-06-20-20170706034741172-00140-00149-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706035435634-00148-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+org,ifaamas)/proceedings/aamas09/pdf/01_full%20papers/02_08_fp_0272.pdf 20170706081902 http://www.ifaamas.org/Proceedings/aamas09/pdf/01_Full%20Papers/02_08_FP_0272.pdf application/pdf 200 GYHX35QJWRJELWJ5GDQZPTPOUUZOCTKF - - 251180 34635154 CITESEERX-CRAWL-2017-06-20-20170706081825105-00419-00428-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706081838210-00420-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+de,fau,cs)/publications/2014/lukas_14_masterthesis.pdf 20170705101722 http://www4.cs.fau.de/Publications/2014/lukas_14_masterthesis.pdf application/pdf 200 GIUQT7SXZ33TWEFBM2MWURJI2M3QE3IW - - 1290532 71068435 CITESEERX-CRAWL-2017-06-20-20170705101605019-00279-00288-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705101714659-00281-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+de,bund,jki,pub)/index.php/jabfq/article/download/3568/4462 20170706041152 http://pub.jki.bund.de/index.php/JABFQ/article/download/3568/4462/ text/html 301 XZBNO24W2ZPQQMJYE6YUUCSRUF7G3ZBT - - 552 417292708 CITESEERX-CRAWL-2017-06-20-20170706040506112-00160-00169-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706041021844-00165-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+whois://whois.arin.net/z+%2B+132.177.133.114 20170713120653 whois://whois.arin.net/z+%2B+132.177.133.114 text/plain - IDEID4YQ6MVJSOE57NPVDLL53ZB3J4DX - - 876 30983517 CITESEERX-CRAWL-2017-06-20-20170707064626094-01007-01015-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170711214025652-01014-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+za,co,csir,researchspace)/dspace/bitstream/10204/4048/1/smith2_2010.pdf 20170706094159 http://researchspace.csir.co.za/dspace/bitstream/10204/4048/1/Smith2_2010.pdf unk 301 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ - - 446 104830407 CITESEERX-CRAWL-2017-06-20-20170706093829986-00509-00518-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706094137978-00512-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+org,annals)/article.aspx?articleid=705034 20170707013120 http://annals.org/article.aspx?articleid=705034 text/html 301 QQYKL57QSERLFM3LXSWMNOFXMOCN7C5G - - 22665 28113974 CITESEERX-CRAWL-2017-06-20-20170707013100780-00967-00976-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170707013100780-00967-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+org,annals)/pdfaccess.ashx?url=/data/journals/aim/20105/0000605-200512200-00013.pdf 20170707045304 http://annals.org/pdfaccess.ashx?url=/data/journals/aim/20105/0000605-200512200-00013.pdf text/html 302 423S7EMGLCVIZ3FLVD7TLAG75HWE4RGI - - 644 222908628 CITESEERX-CRAWL-2017-06-20-20170707042504366-00997-01006-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170707045044604-00999-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+com,sagepub,spi)/content/28/4/501.full.pdf 20170705092027 http://spi.sagepub.com/content/28/4/501.full.pdf unk 301 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ - - 396 553180242 CITESEERX-CRAWL-2017-06-20-20170705091311851-00219-00228-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705091759818-00223-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+ir,mediaj)/favicon.ico 20170705075240 http://mediaj.ir/favicon.ico text/html 404 E3WSNQ7JAFOW7N3ZJ6GLV27T52T25JDK - - 589 455827180 CITESEERX-CRAWL-2017-06-20-20170705074433815-00129-00138-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705075051100-00135-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+com,sagepub,jpr)/content/8/3-4/239.full.pdf 20170705074931 http://jpr.sagepub.com/content/8/3-4/239.full.pdf unk 301 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ - - 400 270368088 CITESEERX-CRAWL-2017-06-20-20170705074433815-00129-00138-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705074843696-00134-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+jp,co,nittuden)/business/pdf/transparent_thermoplastic_resin_with_electron_beam_cross-linking.pdf 20170706083459 http://www.nittuden.co.jp/business/pdf/Transparent_Thermoplastic_Resin_with_Electron_Beam_Cross-Linking.pdf application/pdf 200 V32E3CCO7NMI2M4OHLKG73DXD72LR4B2 - - 715081 761088410 CITESEERX-CRAWL-2017-06-20-20170706082646066-00429-00438-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706083257353-00436-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+lt,lms)/robots.txt 20170705122708 http://www.lms.lt/robots.txt text/plain 200 PF3HTQQT2ULYRWFLJGUWZKHTVZUVMZ2F - - 592 668333707 CITESEERX-CRAWL-2017-06-20-20170705121748408-00399-00408-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705122352502-00406-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+hu,bme,phy)/~szalay/pub/multipartcriteriaposter.pdf 20170705124828 http://www.phy.bme.hu/%7Eszalay/pub/multipartcriteriaPoster.pdf application/pdf 200 L3TUEEZLBJTHAVH74B5N426FAIDBCCOE - - 187866 964760782 CITESEERX-CRAWL-2017-06-20-20170705123641979-00419-00428-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705124315591-00426-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+org,adb,openaccess)/bitstream/handle/11540/1260/new-regime-sme-finance-asia.pdf;jsessionid=f966a3bdac9882ec5a7c326b130f6f81?sequence=1 20170705090940 https://openaccess.adb.org/bitstream/handle/11540/1260/new-regime-sme-finance-asia.pdf%3Bjsessionid%3DF966A3BDAC9882EC5A7C326B130F6F81?sequence%3D1 unk 301 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ - - 515 634039376 CITESEERX-CRAWL-2017-06-20-20170705090333400-00209-00218-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705090728803-00212-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+org,physiology,ajpregu)/content/272/4/r1084 20170706131006 http://ajpregu.physiology.org/content/272/4/R1084 text/html 200 3FOQSKT4WBYOUA6VKKJCEQCN6QF35ANT - - 27346 336293585 CITESEERX-CRAWL-2017-06-20-20170706130432396-00707-00716-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706130850866-00711-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+de,desy,www-it)/common/documentation/cd-docs/sc2002/paperpdf/pap234.pdf 20170705121813 http://www-it.desy.de/common/documentation/cd-docs/SC2002/paperpdf/pap234.pdf application/pdf 200 BONCZ4NNGRNYR22ASFVU7VYTQ24RRNP4 - - 72421 381715704 CITESEERX-CRAWL-2017-06-20-20170705120827801-00389-00398-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705121708700-00397-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+org,oxfordjournals,bmb)/content/28/3/247.full.pdf 20170706014948 http://bmb.oxfordjournals.org/content/28/3/247.full.pdf text/html 301 EJWYVOPONJRARK7SGG6COFRN7CSTHROY - - 643 119398161 CITESEERX-CRAWL-2017-06-20-20170706014800946-00020-00029-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706014907678-00022-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
diff --git a/backfill/tests/test_backfill_hbase_from_cdx.py b/backfill/tests/test_backfill_hbase_from_cdx.py
new file mode 100644
index 0000000..dfed0b3
--- /dev/null
+++ b/backfill/tests/test_backfill_hbase_from_cdx.py
@@ -0,0 +1,55 @@
+
+import io
+import json
+import pytest
+import mrjob
+import happybase_mock
+from backfill_hbase_from_cdx import MRCDXBackfillHBase
+
+@pytest.fixture
+def job():
+ conn = happybase_mock.Connection()
+ conn.create_table('wbgrp-journal-extract-test', {'file': {}, 'grobid0': {}})
+ table = conn.table('wbgrp-journal-extract-test')
+
+ job = MRCDXBackfillHBase(['--no-conf', '-'], hb_table=table)
+ job.hb_table = table
+ return job
+
+#Example to read back rows...
+"""
+def basic_job_run_capturing_output(job):
+
+ job.sandbox(stdin=open('tests/files/example.cdx', 'r'))
+ results = []
+ with job.make_runner() as runner:
+ runner.run()
+ for key, value in job.parse_output(runner.cat_output()):
+ results.append(value)
+
+ print(results)
+ assert len(list(job.hb_table.scan())) == 5
+"""
+
+def test_some_lines(job):
+
+ raw = io.BytesIO(b"""
+com,sagepub,cep)/content/28/9/960.full.pdf 20170705062200 http://cep.sagepub.com/content/28/9/960.full.pdf application/pdf 301 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ - - 401 313356621 CITESEERX-CRAWL-2017-06-20-20170705061647307-00039-00048-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705062052659-00043-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+eu,eui,cadmus)/bitstream/handle/1814/36635/rscas_2015_03.pdf;jsessionid=761393014319a39f40d32ae3eb3a853f?sequence=1 20170705062202 http://cadmus.eui.eu/bitstream/handle/1814/36635/RSCAS_2015_03.pdf%3Bjsessionid%3D761393014319A39F40D32AE3EB3A853F?sequence%3D1 application/PDF 200 MPCXVWMUTRUGFP36SLPHKDLY6NGU4S3J - - 854156 328850624 CITESEERX-CRAWL-2017-06-20-20170705061647307-00039-00048-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705062052659-00043-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+com,pbworks,educ333b)/robots.txt 20170705063311 http://educ333b.pbworks.com/robots.txt text/plain 200 6VAUYENMOU2SK2OWNRPDD6WTQTECGZAD - - 638 398190140 CITESEERX-CRAWL-2017-06-20-20170705062707827-00049-00058-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705063158203-00053-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+""")
+
+ job.sandbox(stdin=raw)
+ job.run_mapper()
+
+ assert job.hb_table.row(b'1') == {}
+ # HTTP 301
+ assert job.hb_table.row(b'3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ') == {}
+ # valid
+ assert job.hb_table.row(b'MPCXVWMUTRUGFP36SLPHKDLY6NGU4S3J') != {}
+ # text/plain
+ assert job.hb_table.row(b'6VAUYENMOU2SK2OWNRPDD6WTQTECGZAD') == {}
+
+ row = job.hb_table.row(b'MPCXVWMUTRUGFP36SLPHKDLY6NGU4S3J')
+ assert row[b'file:mime'] == b"application/pdf"
+ json.loads(row[b'file:cdx'].decode('utf-8'))